Click-through Rate Prediction Combining Feature Importance And Automatic Feature Interaction

In order to cope with the problem of difficult access to information caused by information overload,people are actively looking for solutions,which led to the birth of recommendation systems,and the evolving click-through rate prediction technology became the most important module of recommendation systems.Click-through rate prediction technology based on user behavior sequences can derive vectors that represent users’ interests by analyzing their historical behavior sequences,so as to recommend items that better match their interests to users.The paper investigates techniques related to click-through rate prediction based on feature importance and automatic feature interaction in conjunction with user behavior sequences.Traditional input of user behavior sequence data into a deep neural network for learning introduces noise,loses some information,and cannot accurately represent the user’s interests.In this regard,the paper proposes a click-through rate prediction model combining feature importance based on DIN model,by introducing SENET network to better learn the importance information of features,and combining factorization machine model for low-order feature interaction.In addition,the thesis proposes a click-through rate prediction model combined with automatic feature interaction by introducing a multi-head attention mechanism to achieve highorder interaction of automatically learned features and combining an improved cross network for feature interaction of user information,etc.Using both public and private datasets,we conducted comparative experiments on the model proposed in the paper,designed and implemented a personalized book recommendation system,and proved the effectiveness and practicality of the model proposed in the paper.Specifically,the research work of the dissertation includes the following aspects.(1)To address the problem of how to accurately represent user interest,the paper proposes a click-through rate prediction model combining feature importance,named DISFMN.The model uses a self-attentive mechanism to learn the correlations between items and candidate items in the user’s historical behavior sequence,and assigns different weight information to the features depending on the correlations.The model also introduces the SENET network to learn the weight information of the items themselves in the user’s historical behavior sequence,increasing the weight information of important features while weakening the weight information of unimportant features to reduce the impact from noise.The model also uses a factorization machine model for low-order feature interaction of features to quickly and accurately mine the potential information in the features.The model achieves an accurate representation of user interests with the above modules.(2)To address the problem that feature interactions of all orders cannot be guaranteed to be valid,the paper proposes a click-through rate prediction model combining automatic feature interactions,named DIMHCN.The model uses a self-attention mechanism to acquire user interest while using a multi-head attention mechanism with residual connections to accomplish automatic feature interaction,which can achieve effective high-order feature interaction.The model also uses a modified cross network,using a cross network to interact with high-order features on edge data such as user information and context information as well,to fully explore the hidden information in the data,which can improve the expressiveness of the model compared to the direct Concat operation on user information and context information.(3)The experiments are compared with the classical click-through rate prediction model on public and private datasets respectively,and the experimental effects of the model are tested by evaluation metrics.On all datasets,the experimental results of DISFMN model and DIMHCN model showed significant improvement over the base model,which proved the effectiveness of the model.In addition,an ablation experiment was set up to verify the effect of each module on the effect of the model,and the experimental results showed that each module had a positive effect on the effect of the model,proving the effectiveness of each module.(4)The paper combines the proposed model to design and implement a personalized book recommendation system,which can realize book recommendation for different users’ interests and preferences.In the system implementation part,the book recommendation results match the users’ reading interests,which shows the correctness of the recommendation results and proves the practicality of the proposed model.